I work on LLM inference infrastructure, with a focus on GPU kernels, attention backends, and quantization. My open-source work centers on SGLang, FlashInfer, vLLM, and SageAttention.
| Area | Current work |
|---|---|
| Frameworks | SGLang attention DP routing and load accounting; vLLM attention backend integration, FP8 quantization correctness, DeepGEMM integration, and Model Runner V2 profiling. |
| Kernels | FlashInfer Gated DeltaNet prefill on SM120; SM90 FA3 FP8 paged attention; SageAttention sequence-length handling and paged KV-cache decode for CUDA Graphs. |
| Project | Contribution | Status |
|---|---|---|
| SGLang | Count the unprefilled portion of chunked requests in total_tokens routing load. PR #50088 · Development branch · Reproducible benchmarks. |
Open |
| vLLM | #55732 — SAGE_ATTN backend integration | Open |
| vLLM | #52753 — Pin the static FP8 quantization pattern to group_shape=None |
Open |
| vLLM | #49928 — Keep DeepGEMM FP8 quantization inside an opaque op for TMA scales | Open |
| Project | Contribution | Status |
|---|---|---|
| FlashInfer | #6342 — Experimental SM120 CuTe GDN prefill backend | Merged |
| SageAttention | #402 — SM89/SM120 sequence lengths and paged KV decode for CUDA Graphs | Open |
| sgl-flash-attn | Recover SM90 FA3 FP8 paged hdim256 performance with chunked two-level P@V accumulation while retaining the precision fix. PR #55 · Development branch · Integration into sgl-kernel. |
Merged |
Validation and performance results are documented in the linked PRs and apply to their tested hardware and workloads.

